Ethernet switch architecture can be generally divided into two classes: shared memory and multiprocessor systems. In the shared memory architecture, packets, including a header with a source and destination addresses, a data payload, and a check sum portion at the end of the packet, are received at the ports and then forwarded to a system card via a backplane. Here, the packets are buffered in the same central memory, which is accessed by a single central controller. This controller looks at the destination address of the packet, and possibly other information such as source port and destination port, and then, via an address and other look-up tables, determines what should be done with the packet, such as forward, discard, translate or multicast. This forwarding decision is usually an identification of a LAN Card and a port of that LAN Card in the Ethernet switch that connects to the destination address. The address look-up table tells to which port of the Ethernet switch the packet must be sent to reach the addressed device. The packet is then forwarded to that LAN card. The designated LAN card receives the packet and routes it to the destination port where an Ethernet controller chip sends the packet out on the LAN to the addressed device.
The multiprocessing architecture differs in that a local processor is placed on each one of the cards and each one of these processors accesses and maintains its own address table. As a result, when a multiprocessor-type LAN card receives a packet through one of its ports, it first looks at the destination address and then determines to which one of the other ports on one of the other cards it must be sent and then sends the packet to that card via the backplane connecting the cards. If the port address of the packet happens to be on the same LAN card it was received on, the processor simply sends the packet to that local port address and the packet is never transferred on the backplane. This architecture has certain advantages in that since functionality is replicated between the cards, if any one of the cards should fail the Ethernet switch can still function although this also increases cost. One problem is, however, that a substantial amount of processing and software is devoted to ensuring the address tables on each one of the LAN cards are exact duplicates of each other.
Regarding the handling of the packet within the Ethernet switch, two basic methods are conventional. The first is called store and forward switching. In this switching scheme, a particular LAN card will wait until it has received the entire packet before forwarding it to either another LAN card or the central controller card. This allows the LAN card to confirm the packet is valid and uncorrupted by reference to the check sum information contained at the end of the packet. A new approach has been proposed which is called cut-through switching, a purpose of which is to decrease packet latency. Here, not yet fully received packets are forwarded to the destination port and begun to be sent out or broadcasted from the switch before the entire packet has been received. This both decreases packet latency and also decreases the amount of buffering RAM required by each LAN card. The problem with this approach, however, is that if the packet turns out to be invalid or corrupted there is no way to drop the packet since it is already been started to its destination.